About the inference time

[Jerryme-xxm]

Thanks for your great work!

These days I try to compress MobileVit, and reduce the FLOPs and parameters, but the inference time almost doesn't change.

Only when I remove some MV2 blocks, the inference time reduce a little.

My method I used for compressing:
1: Try LinFormer.(This method reduces the FLOPs, but increases the length of sequence. It seems longer sequence means slower speed)
2: Decrease the channels of some layers.
3: Other methods, such as replacing conv33 with conv11(learning from you)

Could you please tell me the reason?

[Jerryme-xxm]

Just now, I test a single conv11 and conv33.
conv11: FLOPs 49152 time 0.0471
conv33: FLOPs 442368 time 0.0520
Sad...